Normative Competence Challenges in Multi-Agent LLM Environments

LLM agents fail to learn community-specific norms through interaction alone, even when such learning improves accuracy. The study uses a multi-agent community debate setting with synthetic norms to isolate normative competence from static pretraining knowledge.